Are you really into object recognition, but you are sick of looking at 2D boxes and 2D masks? Let's play with 3D shapes!
We build on Mask R-CNN and extend it to infer 3D meshes. Given an input image, we detect all objects, infer their 2D instance boxes and masks as well as their 3D object shapes, all end-to-end! Naturally, we call our approach Mesh R-CNN :D
Paper: https://arxiv.org/abs/1906.02739
Joint work with Jitendra Malik and Justin Johnson
P.S. This task is really hard!
